Ask a chatbot for a week of meals and it will hand you one: breakfasts, lunches, dinners, a shopping list, calorie figures to one decimal place. The format is convincing enough that the obvious question — is any of this right? — tends to arrive after the plan has already been followed for a fortnight.
October 2026 produced an unusually direct set of answers, and they do not point in a single direction. One study had dietitians blind-rate AI-assisted meal plans against plans written by dietitians, and the machine side scored higher. Three other papers took the same kind of output and recalculated the nutrients independently, and the numbers came apart.
The blinded test, and who was holding the pen
The comparison, published 16 October in Nutrition, ran as a cross-sectional web study. Fifty-four eligible anonymous evaluator records from Brazilian dietitians contributed 130 evaluations of five matched plan pairs; each pair held one plan from a constrained AI-assisted pipeline and one designed by a dietitian for the same fictional persona. Evaluators rated nine items across three domains, with an optional side-by-side preference task. An overall score, which required all nine items, was available for 124 evaluations.
The pipeline’s plans drew a mean rating of 3.67 (standard deviation 0.57, n = 59) against 3.11 (0.73, n = 65) for the dietitian-designed comparators, a difference the authors put at d = 0.86. Among the 38 evaluator records with complete scores for both origins, 31 favoured the pipeline, six favoured the dietitian, and one scored them equally. The gap held in all five persona pairs and in 32 of the 42 optional pairwise choices. The authors describe 3.67 as sitting above the neutral anchor of the five-point scale rather than near the top of it.
Two details do most of the interpretive work here. The first is the phrase “constrained pipeline”: this was a structured system with specified inputs, not a bare chatbot answering a typed question. The second is on the disclosure line. All of the authors are named on the Brazilian software registration for NextNutri, the tool behind the pipeline, listed as a current non-financial interest and a potential future financial one.
The study also measured something narrower than nutrition. Dietitians judged plans on paper, for invented personas; nobody ate the food, and no one tracked weight, blood markers or how closely anyone stuck to the plan. The authors’ own conclusion is correspondingly modest: the results “support prospective evaluation of a structured first draft workflow, with required clinician edits and patient-relevant outcomes as endpoints.” A first draft that a clinician edits is a different product from an app that generates a plan and sends it.
Then the nutrients were recalculated
The counterweight comes from evaluations that did not ask whether a plan looked sensible but whether its arithmetic held.
In Nutrients on 28 September, researchers in Türkiye took anonymised records from six women with heterogeneous polyendocrine metabolic ovarian syndrome (the framing the paper uses for PCOS), turned them into standardised prompts, and put them to ChatGPT-4, Claude Sonnet 4.6, Gemini 2.5 Flash and Mistral 2.0 through their free web interfaces. Three generations per case and model produced 216 daily menus. The models’ own nutrient figures were then compared with calculations from BeBiS 9.0, a professional nutrition-analysis package, and two experts rated the first plan from each combination against a clinical rubric and a red-flag checklist.
No model was consistently accurate. The best performer shifted by nutrient — Claude Sonnet 4.6 for energy, carbohydrate and fibre, ChatGPT-4 for protein, Mistral 2.0 for fat — while Gemini 2.5 Flash showed the largest percentage errors across the board and a mean energy bias of +531.1 kcal per day. Reproducibility was poor for most models; ChatGPT-4 was comparatively steadier. Rankings on clinical appropriateness did not survive correction for multiple comparisons, and Gemini produced the most confirmed red flags (28), with all six cases tripping the study’s “masked low-energy” indicator. The authors’ conclusion is blunt: the findings “do not support their independent or unsupervised use as individualized clinical nutrition recommendations.”
A second Nutrients paper, published 25 September, ran the same kind of check in a condition where arithmetic is the treatment. Eight standardised simulated cases of classical phenylketonuria spanned infancy, adolescence, adulthood and maternal PKU; ChatGPT-5.2, Claude 4.5 and Gemini 3 produced 48 plans across two prompting conditions, recalculated in a separate nutrition program and scored by three blinded paediatric metabolic dietitians. Phenylalanine provision came out below the case-specific prescription in 36 of the 48 plans — 75.0%. Individual nutrient comparisons between what the models reported and what independent calculation found did not separate statistically after correction, but plan-level error varied widely, and the authors note that numerical plausibility does not establish agreement with a clinical prescription.
Fluency fails in smaller ways too. A September audit in Frontiers in Nutrition put 15 consumer-style prompts to five model configurations, ten times each, for a standardised recreational female marathon runner — 750 responses in total. Carbohydrate advice for one race-fueling scenario ranged from 15 to 120 g/h in one model and 5 to 145 g/h in another, against the 30–90 g/h comparison range the authors used. One response set a target of 115–145 g/h while its own gel schedule delivered 37.5–45 g/h. The advice read smoothly in every case.
What the better-performing setups had in common
Neither the enthusiasm nor the errors are properties of “AI” in the abstract; they track how the system is built around the model. In the phenylketonuria study, prompts carrying full clinical detail drew higher expert ratings than minimal zero-shot prompts. A September in-silico study in the Journal of Medical Internet Research makes the same point from the other side: a guideline-integrated, safety-aware framework for kidney-stone dietary advice beat a standard model configuration across metabolic specificity, guideline adherence and actionability (median 5.00 versus 3.00 on the first, all comparisons P < .001), with safety pass rates of 30/30 against 25/30. When the same cases were rephrased the way a patient would type them, the advantage narrowed. That was 30 synthetic vignettes, not patients, and the framework was the authors’ own — but it locates the gain where clinicians would expect it, in the explicit rules and checks wrapped around the language model.
Where a draft stops being low-stakes
For a generally healthy adult who wants ideas for lunches, a generated plan is a starting point that costs nothing and can be checked against ordinary guidance. MedlinePlus is careful about what that guidance can say in advance: “Good nutrition isn’t one-size-fits all. There’s no single diet or calorie amount that works for everyone. Your needs depend on your age, sex, activity level, health conditions, and health goals.” The same page suggests asking a health care provider about a diet that fits, and notes that they may refer you to a dietitian for specialised recommendations. The current Dietary Guidelines for Americans make the same move in the other direction, leading with whole, healthy and nutritious foods rather than a plan document.
The stakes change with the condition. A diet built around a metabolic disorder, advanced kidney disease, pregnancy, a food allergy or insulin-treated diabetes is a prescription with numbers attached, and the papers above are explicit that model output is not a substitute for the clinician’s calculation. Anyone in that position, or anyone who has had an eating disorder, is not the use case a chatbot’s meal plan was built for.
The narrow reading of this month’s evidence is not that dietitians should be replaced, nor that AI has nothing to offer. It is that the two kinds of test measure different things and both results are real: blinded professionals judged the machine-assisted drafts at least as usable as the ones written by hand, and independent recalculation found the numbers unreliable enough that the researchers who ran those audits advised against unsupervised use. A draft can be genuinely good and still need an editor. Anyone using one for anything beyond inspiration should treat it that way — and, as our medical disclaimer sets out, bring persistent or condition-specific questions to a healthcare professional rather than to a prompt. Our earlier coverage of the professional consensus on weight-loss diets makes a related point: even where diets converge, the individual is where the decisions get made.